Learning method for neural network having discrete interconnection strengths

This application is a continuation application Ser. No. 07/953,083, filed Oct. 19, 1992, now abandoned which is a continuation in part of application Ser. No. 07/622,516, filed Dec. 5, 1990 now abandoned.

This application claims the benefit under 35 U.S.C. §120 and is a continuing application of application of the same title, Ser. No. 07/622,516 filed Dec. 5, 1990, now abandoned.

1. Field of the Invention

The present invention relates to a learning method in a neural network having discrete interconnection strengths, such as an optical neuron computer.

2. Description of the Related Art

FIG. 1 illustrates the structure of a typical layered neural network. FIG. 2 is a flow chart which illustrates the operation of such a neural network. A layered network of the type described above has been disclosed, for example, in Primer on Neural Networks, pp. 55 to 57, edited by Amari and Mori, published by Trikeps Corporation, Nov. 30, 1985.

Referring to FIG. 1, reference numerals 1a, 1b, represent neurons which constitute an input layer of a neural network 2b, 2b , . . . , represent neurons which constitute an intermediate layer; 3a, 3b , . . . , represent neurons which constitute an output layer; 4a, 4b , . . . , represent inter-connections between the neurons in the input layer and those in the intermediate layer; 5a, 5b , . . . , represent inter-connections between neurons in the intermediate layer and those in the output layer 6a, 6b , . . . , represent input signals; and 7a, 7b , . . . , represent output signals.

In accordance with a total input I and threshold θ, a single neuron transmits an output V given from the following equation:

V=f(I-θ) where f(I-θ)=1/[1+exp (-I- θ)]. (1)

Allowing the number of neurons in the input layer, the intermediate layer and the output layers to be n I , n H and n O , respectively, the output from the i-th neuron in each of the layers to be V I i, V H i and V O i, respectively, and the thresholds to be θ I i, θ H i and θ O i, respectively, the outputs from the i-th neurons in the intermediate layer and the output layer can be respectively given from the following equations: ##EQU1## where W IH ij represents the interconnection strength between the i-th (i=1, 2, . . . , n H ) neuron in the intermediate layer and the j-th (j=1, 2, . . . , n I ) neuron in the input layer and W HO ij represents the interconnection strength between i-th (i=1, 2, . . . , n O ) neuron in the output layer and the j-th (j=1, 2, . . . , n H ) neuron in the intermediate layer.

Therefore, when an output V I =(V I 1, V I 2 . . . , V I n I is determined, in response to an input signal to the neurons 1a, 1b, . . . , which constitute the input layer, an output V O= (V O 1, V O 2 . . . , V O n O ) from the neurons 3a, 3b, . . . , , which constitute the output layer, is determined from Equations (2) and (3).

Usually, "learning" in a neural network is a process of updating the interconnection strengths W IH ij and W HO ij in such a manner that the neurons 3a, 3b, . . . , in the output layer transmit a specific output V O (m) with respect to a certain output V I (m) from the neurons 1a, 1b, . . . , in the input layer. In the learning process, the above-described plurality of specific outputs V I (m) from the input layer and corresponding desired outputs V O (m) from the output layer are called "educator signals".

The operation of learning in such a neural network will now be described with reference to the flowchart of FIG. 2.

First, all of the initial values of the interconnection strengths are set by using random numbers (step ST1). The variable m (m=1, 2, . . . , M), for denoting the m-th educator signal is initialized (step ST2). Then, the actual output signal V O from the neural network, corresponding to V I (m) of the m-th educator signal, is calculated by using Equations (2) and (3) (step ST3). A quantity for updating each interconnection strength is then calculated by using the following equations (step ST4): ##EQU2##

ΔW.sup.HO ij=-αV.sup.O i(1-V.sup.O i)V.sup.H j(V.sup.O i-V.sup.O(m) i) (5)

where ΔW IH ij and ΔW HO ij respectively represent the update quantity by which interconnection strengths W IH ij and W HO ij are updated and where coefficient α(0<α≦1) represents the learning rate.

Then, each interconnection strength is updated according to the following equation by the update quantity obtained in step ST4 (step ST5):

W.sup.IH ij(new)=W.sup.IH ij(old)+ΔW.sup.IH ij (6)

W.sup.HO ij(new)=W.sup.HO ij(old)+ΔW.sup.HO ij (7)

For each educator signal m (M signals), the square error between the actual output signal V O hand the desired output V O (m), corresponding to V I (m) of the educator signal m, is calculated for each neuron in the output layer. If any of the sums of these square errors is less than a convergence criterion value ε (usually a value less than 0.01 n O is used), learning is ended (step ST6). If any of the above-described square errors exceeds ε and the educator signal which has been learned immediately before is the M-th signal (step ST7), the variable m is again initialized (step ST2). Then learning is repeated by incrementing the variable m by one until the actual outputs V O of the network, with respect to all V I (m) of the M educator signals are sufficiently approximate to all of the desired outputs V O (m) (step ST8).

In the conventional learning method for a neural network, the inter-connection strengths must be taken from a continuous range of values in order to be updated. Therefore, learning cannot be performed using this method in a neural network which has a discrete interconnection strength, such as in an optical neuro-computer.

Accordingly, an object of the present invention is to provide a learning method for use in a neural network having a discrete interconnection strength and to improve the function and performance of such a neural network.

The present invention provides a learning method for a neural network, having an interconnection strength which is discrete. The method of learning comprises the steps of calculating a quantity of update based on the discrete interconnection strength, updating an imaginary interconnection strength, taken from a continuous range of values, by the calculated quantity of update, and discretizing the updated imaginary interconnection strength so as to modify and update the interconnection strength.

The learning method in a neural network according to the present invention is such that an imaginary interconnection strength, taken from a range of continuous values, is induced. A quantity of update, calculated using a discrete interconnection strength related to the imaginary interconnection strength, is added to the imaginary interconnection strength, which are then discretized. Therefore, the quantities of update of the same interconnection strengths are stored in the imaginary interconnection strength. Therefore, learning in a neural network having a discrete interconnection strength can be performed. As a result, the function and performance of the neural network can be improved.

Other and further objects, features and advantages of the invention will be more apparent from the following description.

FIG. 1 illustrates the structure of a typical layered neural network;

FIG. 2 is a flowchart which illustrates the operation of a conventional learning method for a neural network;

FIG. 3 is a flowchart which illustrates the operation of a learning method for a neural network according to an embodiment of the present invention;

FIGS. 4A and 4B illustrate the intensity distribution of imaginary interconnection strength induced according to the present invention; and

FIG. 5 illustrates a computer system suitable for implementing the learning method of the present invention.

A preferred embodiment of the present invention will now be described with reference to the drawing. FIG. 3 is a flowchart which illustrates the operation of a neural network according to the present invention in which the back-propagation method is applied to a neural network having discrete interconnection strengths. The neural network is generally similar to a conventional neural network (see FIG. 1), in the form of a triple-layer shape which learns educator signals V I (m), V O (m) (m=1, 2, . . . , M). However, the interconnection strength of interconnections 4a and 4b, between the neurons in the input layer and the neurons in the intermediate layer, and the interconnection strength of interconnections 5a and 5b, between the neurons in the intermediate layer and neurons in the output layer, are limited to discrete values (for example, 0, 1 and 2). Such neural networks include optical and opto-electronic neural networks, optical neuro-computers, resistor-type neural networks, digital computer neural network implementations, and neural networks using pulse-stream arithmetic.

FIGS. 4A and 4B illustrate the relationship of imaginary interconnection strengths to discrete interconnection strengths. FIG. 4A illustrates the distribution of imaginary interconnection strengths W IH ij and W HO ij and FIG. 4B illustrates the distribution of discrete strengths W IH ij and W HO ij. The discrete strength is -W when the imaginary interconnection strength which takes a continuous value is less than -W/2; 0 when the imaginary interconnection strength is greater than -W/2 and less than W/2; and W when the imaginary interconnection strength is greater than W/2.

The operation of a learning method with this neural network will now be described with reference to the flowchart of FIG. 3 and system diagram of FIG. 5.

The system of FIG. 5, suitable for practicing the learning method of the invention, includes a computer 20 having a processor 22, memory 24 and Input/Output (I/O) controller 26. The processor 22 performs arithmetical and logical operations and controls the sequence of such operations in a manner well-known in the art.

The memory 24 stores: the set of educator signals (m pairs of V I (m) inputs and corresponding desired V O (m) outputs); information concerning neuron states (i.e., thresholds θ) from the neural network; imaginary weights taken from a range of continuous values; and the corresponding discretized (quantized) weights for the neural network.

The I/O controller 26 connects computer 20 to a neural network 28 which has discrete weights. Input signals and the quantized weights are applied to the neural network 28 by the computer 20 through this controller 26.

Alternatively, the set of discrete weights need not be stored in the quantized weight portion of memory 24 where the neural network performs discretization, or quantization, directly. The imaginary weights can be applied directly to the neural network which in turn quantizes the imaginary weights to obtain the discrete weights. For example, a simple threshold circuit in the neural network 28 may provide this function of quantization.

In opto-electronic neural networks, including optical neuron computers, the interconnection strengths are provided by a spatial light modulator (SLM). In an SLM, an array of pixels takes on binary values (i.e., a pixel is either transparent or opaque).

In resistor-type neural networks, an array of amplifiers have their inputs and outputs interconnected through a matrix of resistive coupling elements. The coupling elements are programmable to be either on or off. (See Graf, H. P. et al., "VLSI Implementation of a Neural Network Model", Computer, March 1988, p. 41). Alternatively, outputs of amplifiers may be connected to control two switches. Between two amplifiers, two memory cells control two other switches, respectively. Three connection states are possible: excitatory, inhibitory and disabled.

A neural network with discrete interconnection strengths may also be implemented using digital programmable elements as shown in Duranton, U.S. Pat. No. 4,994,982. This method may be applied to such a neural network as well.

Another type of neural network on which this learning method may be used is a VLSI neural network using pulse-stream arithmetic, such as shown in Murray, A. F., "Asynchronous VLSI Neural Networks Using Pulse-stream Arithmetic," IEEE Journal of Solid-State Circuits, Vol. 23, No. 3, June 1988. Such a network has digitally programmable weights. In the cited reference, the weight is 5 digital bits, corresponding to a discrete range of 32 weights.

Referring now to FIG. 3, the imaginary interconnection strengths W IH ij and W HO ij are initialized in the memory 24 as random numbers, according to the distribution shown in FIG. 4A (step ST9). The imaginary interconnection strengths are then discretized and stored in memory 24 so as to determine the actual interconnection strengths W IH ij and W O ij (initial values) of the neural network (step ST10). FIG. 4B illustrates a case where there are three possible discrete values for the interconnections. However, the present invention is not limited to this arrangement. The range of possible discrete values depends on the actual neural network used.

Next, the variable m (m=1, 2, . . . , M), denoting the number of educator signals, is initialized (step ST11). After the interconnection strengths in the neural network 28 are set to the discretized values, the actual output signal V O from the neural network in response to V I (m) of the m-th input educator signal is obtained. The actual output can be defined by using the discrete interconnection strengths W IH ij and W HO ij in accordance with Equations (2) and (3) (step ST12).

Then, the quantities of update ΔW IH ij and ΔW HO ij are calculated using equations (4) and (5) by using the discrete strengths W IH ij and W HO ij (step ST13). The calculated quantity of update is added to the imaginary interconnection strengths W IH ij and W HO ij using equations (6) and (7) (step ST14) to update them. Then, the imaginary interconnection strengths W IH ij and W HO ij are discretized in accordance with the relationship defined by FIGS. 4A and 4B. Thus, updated discrete interconnection strengths W IH ij and W HO ij of the neural network are determined (step ST15).

For all of the educational signals (M signals), the square error between the actual output signal V O and the desired output signal V O (m) with respect to V I (m) of the educator signal, is calculated for each neuron of the output layer. If any of the sums of the square errors is less than a convergence criterion value ε, learning is ended (step ST16). If any of the above-described square errors exceeds ε, and the educator signal which has been learned immediately before is the M-th signal (step ST17), the variable m is again initialized (step ST11). Learning is then repeated by incrementing the variable m by one until the actual output V O of the neural network in response to V I (m) is suitably approximate to the corresponding desired output V O (m) (correct output), with respect to all of the educator signals.

As described above, the quantities of update ΔW IH ij and ΔW HO ij are calculated by using the discrete interconnection strengths, W IH ij and W HO i.sup. are added to the imaginary interconnection strengths, W IH ij and W HO ij, which take continuous values. The imaginary interconnection strength is then discretized. As a result, an updated discrete interconnection strength is determined. Therefore, even if the quantities of a sole update operation, ΔW IH ij and ΔW HO ij, are smaller than the difference between the values of the discrete interconnection strengths of the neural network, the above-described quantities of update are added to the imaginary interconnection strength, and are thus stored. As a result, the interconnection strength of the network can be varied between the discrete values after being subjected to a plurality of updating operations. As a result, the interconnection strength can be updated and learning can be performed in a neural network having a discrete interconnection strength.

According to this embodiment, the back-propagation method is applied to a neural network having a discrete interconnection strength. The present invention may be applied to other learning methods. A similar effect can be obtained by improving other learning methods in such a manner that an imaginary interconnection strength taking continuous values is induced, the imaginary interconnection strength is updated by using a quantity of update obtained by using the discrete interconnection strength of the neural network, and the imaginary interconnection strength is discretized to update the discrete interconnection strength of the neural network.

Although the preferred embodiment of the invention has been described in detail it is to be understood that the details of construction and the combination and arrangement of parts in the present disclosure of the preferred embodiment may be changed without departing from the spirit and scope of the invention as hereinafter claimed.